Jeremy Burton, Snowflake's general manager of observability, argues managing critical business data has never been a commodity, rejecting frontier model companies' view of the data layer.
Burton says Observe built on Snowflake rather than its own database, trading part of its engineering team for effort spent on value above the database.
Burton says large enterprises now generate 500 terabytes to a petabyte of telemetry a day, a scale that a language model cannot simply be pointed at.
Why the data layer, not the model, decides enterprise AI
Observability is a data problem, and Snowflake's bet is that the data layer decides who wins in enterprise AI. Telemetry has grown from petabyte rarities to petabytes a day, and agents fed raw or conflicting data give confident wrong answers. Companies that want a return on AI have to build the data foundation first.
Takeaways
Observability is fundamentally a data problem. Teams run a separate tool for each kind of telemetry, which makes monitoring confusing and expensive. Burton argues the fix sits in the data layer, where all telemetry can live together, and the problem is still much as it was seven years ago.
Telemetry volumes now dwarf anything from a decade ago. When Burton joined EMC about fifteen years ago, one customer in the world had a petabyte. Midsized firms now produce 30 to 50 terabytes of telemetry daily, large enterprises 500 terabytes to a petabyte, and some collect a petabyte every two days.
Frontier labs are wrong that data is a commodity. Burton says model companies, including Anthropic, treat the data layer as a commodity. He counters that managing critical business data never was one, and an agent given ten sources of truth may average them into a confidently wrong answer.
Enterprise AI returns depend on the data foundation. A language model cannot simply be pointed at hundreds of terabytes of telemetry, and teams that try burn tokens and blow budgets. More context reduces token burn, and a bank's app knowledge is not on the public internet.
Knowledge graphs let agents do the legwork. Observe built a knowledge graph in 2018 that AI agents can now build and maintain. Burton wants agents doing 90 percent or more of the work, while implementation staff sit with customers' SRE teams to learn their troubleshooting questions.
LLMs need their own observability built on evaluations. Burton calls LLMs a new animal in the zoo, nondeterministic and less observable than databases, so evaluations will become core to monitoring them. He expects telemetry to grow sharply again, as it did when Kubernetes arrived.
AI changes jobs and raises the value of engineering. Burton uses the tractor analogy to argue AI will change work rather than eliminate it. With implementation becoming cheap, software engineering and architecture practice matter more, and he stays bullish on software engineers.
How Observe bet on Snowflake
Burton joined Observe, a startup with a handful of founding engineers, because after a career at big companies he wanted to build something from scratch.
He talked to three or four companies building on top of Snowflake before choosing Observe.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Observe chose to build on Snowflake rather than its own database, a decision that cost the company part of its engineering team.
The bet was that engineers could then spend their effort on value above the database.
Within six months the team proved an interactive observability experience could run on Snowflake.
Snowflake later acquired Observe, and Burton now runs observability there as a general manager.